Implement orchestration logic for agent coordination

Build the main orchestration loop that spawns agents and manages handoffs.

Flow: 1. Initialize shared workspace with question 2. Spawn Planner agent, pass workspace 3. Wait for Planner to complete, update workspace 4. Spawn Executor agent, pass updated workspace 5. Wait for Executor to complete, update workspace 6. Spawn Verifier agent, pass updated workspace 7. Check Verifier decision:

• If APPROVED: return final answer
• If NEEDS_REVISION and iterations < max: goto step 2 or 4 (depending on what needs fixing)
• If max iterations reached: return current answer candidate

Requirements:

• Use existing spawn_subagent infrastructure
• Handle agent failures gracefully
• Track total cost/tokens across all agents
• Return final workspace state along with answer for debugging
• Timeout: max 10 minutes total runtime

Integration point:

• Add a new tool or function callable from Ava: solve_gaia_question(question: str) -> dict